Especially when compiling -DCLIB_VEC64=1, the code performs up to three orders of magnitude better when we avoid creating vectors